Machine learning for in vivo soft tissue strain prediction
Measuring in vivo strain is challenging. Previously, researchers have measured strain in vivo by combining medical imaging with classical image texture correlation algorithms, e.g., direct deformation estimation (DDE) and digital image correlation (DIC). However, these techniques are often limited by low signal-to-noise ratios, poor image texture, and other image artifacts. Thus, our lab has developed a deep convolutional neural network, StrainNet, to automatically measure patient-specific in vivo strains from high-frequency ultrasound images (Fig. A). Applying StrainNet to ultrasound images of human flexor tendons (Fig. B), we have been able automatically to segment the tendon as well as accurately capture strain distributions (Fig. C) by passing image sequences through a deep convolutional neural network (Fig. D). Currently, StrainNet has demonstrated a greatly improved predictive power with a mean strain error 90% lower than DDE and DIC (Fig. E). We believe StrainNet will generalize to other biological domains beyond soft tissues as we aim to uncover in vivo tissue behavior under loading, in order to better understand tissue injury and rehabilitation mechanisms.
If you are interested in applying machine learning to medical images, please feel free to contact Reece Huff (rdhuff at berkeley dot edu).